result_folder = './Results'
if not os.path.isdir(result_folder):
os.mkdir(result_folder)
if __name__ == '__main__':
data=pd.DataFrame(columns=['image_name', 'word_bboxes', 'pred_words', 'align_text'])
data['image_name'] = image_names
# load net
net = CRAFT() # initialize
print('Loading weights from checkpoint (' + args.trained_model + ')')
if args.cuda:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model)))
else:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model, map_location='cpu')))
if args.cuda:
net = net.cuda()
net = torch.nn.DataParallel(net)
cudnn.benchmark = False
net.eval()
# LinkRefiner
refine_net = None
if args.refine:
from refinenet import RefineNet
refine_net = RefineNet()
def ground_truth(args):
# initiate pretrained network
net = CRAFT() # initialize
print('Loading weights from checkpoint (' + args.trained_model + ')')
if args.cuda:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model)))
else:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model, map_location='cpu')))
if args.cuda:
net = net.cuda()
net = torch.nn.DataParallel(net)
cudnn.benchmark = False
net.eval()
filelist, _, _ = file_utils.list_files('/home/ubuntu/Kyumin/Autotation/data/IC13/images')
for img_name in filelist:
# get datapath
if 'train' in img_name:
label_name = img_name.replace('images/train/', 'labels/train/gt_').replace('jpg', 'txt')
else:
label_name = img_name.replace('images/test/', 'labels/test/gt_').replace('jpg', 'txt')
label_dir = img_name.replace('Autotation', 'craft').replace('images', 'labels').replace('.jpg', '/')
os.makedirs(label_dir, exist_ok=True)
image = imgproc.loadImage(img_name)
gt_boxes = []
gt_words = []
with open(label_name, 'r', encoding='utf-8-sig') as f:
lines = f.readlines()
for line in lines:
if 'IC13' in img_name: # IC13
gt_box, gt_word, _ = line.split('"')
if 'train' in img_name:
x1, y1, x2, y2 = [int(a) for a in gt_box.strip().split(' ')]
else:
x1, y1, x2, y2 = [int(a.strip()) for a in gt_box.split(',') if a.strip().isdigit()]
gt_boxes.append(np.array([[x1, y1], [x2, y1], [x2, y2], [x1, y2]]))
gt_words.append(gt_word)
elif 'IC15' in img_name:
gt_data = line.strip().split(',')
gt_box = gt_data[:8]
if len(gt_data) > 9:
gt_word = ','.join(gt_data[8:])
else:
gt_word = gt_data[-1]
gt_box = [int(a) for a in gt_box]
gt_box = np.reshape(np.array(gt_box), (4, 2))
gt_boxes.append(gt_box)
gt_words.append(gt_word)
score_region, score_link, conf_map = generate_gt(net, image, gt_boxes, gt_words, args)
torch.save(score_region, label_dir + 'region.pt')
torch.save(score_link, label_dir + 'link.pt')
torch.save(conf_map, label_dir + 'conf.pt')
use_cuda = torch.cuda.is_available()
device = 'cuda:0' if use_cuda else 'cpu'
print('Load the synthetic data ...')
data_loader = Synth80k('D:/Datasets/SynthText')
train_loader = torch.utils.data.DataLoader(data_loader,
batch_size=1,
shuffle=True,
num_workers=0,
drop_last=True,
pin_memory=True)
batch_syn = iter(train_loader)
print('Prepare the net ...')
net = CRAFT()
net.load_state_dict(copyStateDict(
torch.load('./weigths/synweights/0.pth')))
net.to(device)
data_parallel = False
if torch.cuda.device_count() > 1:
net = nn.DataParallel(net)
data_parallel = True
cudnn.benchmark = False
print('Load the real data')
real_data = ICDAR2013(net, 'D:/Datasets/ICDAR_2013')
real_data_loader = torch.utils.data.DataLoader(real_data,
batch_size=5,
shuffle=True,
num_workers=0,
drop_last=True,
pin_memory=True)
torch.save(score_region, label_dir + 'region.pt')
torch.save(score_link, label_dir + 'link.pt')
torch.save(conf_map, label_dir + 'conf.pt')
if __name__ == '__main__':
import ocr
score_region = torch.load('/home/ubuntu/Kyumin/craft/data/IC13/labels/train/100/region.pt')
score_link = torch.load('/home/ubuntu/Kyumin/craft/data/IC13/labels/train/100/link.pt')
conf_map = torch.load('/home/ubuntu/Kyumin/craft/data/IC13/labels/train/100/conf.pt')
image = imgproc.loadImage('/home/ubuntu/Kyumin/Autotation/data/IC13/images/train/100.jpg')
print(score_region.shape, score_link.shape, conf_map.shape)
# cv2.imshow('original', image)
cv2.imshow('region', imgproc.cvt2HeatmapImg(score_region))
cv2.imshow('link', score_link)
cv2.imshow('conf', conf_map)
net = CRAFT().cuda()
net.load_state_dict(test.copyStateDict(torch.load('weights/craft_mlt_25k.pth')))
net.eval()
_, _, ref_text, ref_link, _ = test.test_net(net, image, ocr.argument_parser().parse_args())
cv2.imshow('ref text', imgproc.cvt2HeatmapImg(ref_text))
cv2.imshow('ref link', ref_link)
cv2.waitKey(0)
cv2.destroyAllWindows()
def train(args):
# load net
net = CRAFT() # initialize
if not os.path.exists(args.trained_model):
args.trained_model = None
if args.trained_model is not None:
print('Loading weights from checkpoint (' + args.trained_model + ')')
if args.cuda:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model)))
else:
net.load_state_dict(test.copyStateDict(torch.load(args.trained_model, map_location='cpu')))
if args.cuda:
net = net.cuda()
net = torch.nn.DataParallel(net)
cudnn.benchmark = False
# # LinkRefiner
# refine_net = None
# if args.refine:
# from refinenet import RefineNet
#
# refine_net = RefineNet()
# print('Loading weights of refiner from checkpoint (' + args.refiner_model + ')')
# if args.cuda:
# refine_net.load_state_dict(test.copyStateDict(torch.load(args.refiner_model)))
# refine_net = refine_net.cuda()
# refine_net = torch.nn.DataParallel(refine_net)
# else:
# refine_net.load_state_dict(test.copyStateDict(torch.load(args.refiner_model, map_location='cpu')))
#
# args.poly = True
criterion = craft_utils.CRAFTLoss()
optimizer = optim.Adam(net.parameters(), args.learning_rate)
train_data = CRAFTDataset(args)
dataloader = DataLoader(dataset=train_data, batch_size=args.batch_size, shuffle=True)
t0 = time.time()
for epoch in range(args.max_epoch):
pbar = tqdm(enumerate(dataloader), total=len(dataloader), desc=f'Epoch {epoch}')
running_loss = 0.0
for i, data in pbar:
x, y_region, y_link, y_conf = data
x = x.cuda()
y_region = y_region.cuda()
y_link = y_link.cuda()
y_conf = y_conf.cuda()
optimizer.zero_grad()
y, feature = net(x)
score_text = y[:, :, :, 0]
score_link = y[:, :, :, 1]
L = criterion(score_text, score_link, y_region, y_link, y_conf)
L.backward()
optimizer.step()
running_loss += L.data.item()
if i % 2000 == 1999 or i == len(dataloader) - 1:
pbar.set_postfix_str('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / min(i + 1, 2000)))
running_loss = 0.0
# Save trained model
torch.save(net.state_dict(), args.weight)
print(f'training finished\n {time.time() - t0} spent for {args.max_epoch} epochs')